* [PATCH] gcc: xtensa: reorder movsi_internal patterns for better code generation during LRA
@ 2024-03-14 17:30 Max Filippov
0 siblings, 0 replies; only message in thread
From: Max Filippov @ 2024-03-14 17:30 UTC (permalink / raw)
To: gcc-patches; +Cc: Takayuki 'January June' Suwa, Max Filippov
After switching to LRA xtensa backend generates the following code for
saving/loading registers:
movi a9, 0x190
add a9, a9, sp
s32i.n a3, a9, 0
instead of the shorter and more efficient
s32i a3, a9, 0x190
E.g. the following code can be used to reproduce it:
int f1(int a, int b, int c, int d, int e, int f, int *p);
int f2(int a, int b, int c, int d, int e, int f, int *p);
int f3(int a, int b, int c, int d, int e, int f, int *p);
int foo(int a, int b, int c, int d, int e, int f)
{
int g[100];
return
f1(a, b, c, d, e, f, g) +
f2(a, b, c, d, e, f, g) +
f3(a, b, c, d, e, f, g);
}
This happens in the LRA pass because s32i.n and l32i.n are listed before
the s32i and l32i in the movsi_internal pattern and alternative
consideration loop stops early.
gcc/
* config/xtensa/xtensa.md (movsi_internal): Move l32i and s32i
patterns ahead of the l32i.n and s32i.n.
---
gcc/config/xtensa/xtensa.md | 12 ++++++------
1 file changed, 6 insertions(+), 6 deletions(-)
diff --git a/gcc/config/xtensa/xtensa.md b/gcc/config/xtensa/xtensa.md
index 1a2249b059a0..5cdf4dffe700 100644
--- a/gcc/config/xtensa/xtensa.md
+++ b/gcc/config/xtensa/xtensa.md
@@ -1270,13 +1270,15 @@
})
(define_insn "movsi_internal"
- [(set (match_operand:SI 0 "nonimmed_operand" "=D,D,D,D,R,R,a,q,a,a,W,a,a,U,*a,*A")
- (match_operand:SI 1 "move_operand" "M,D,d,R,D,d,r,r,I,Y,i,T,U,r,*A,*r"))]
+ [(set (match_operand:SI 0 "nonimmed_operand" "=D,D,D,a,U,D,R,R,a,q,a,a,W,a,*a,*A")
+ (match_operand:SI 1 "move_operand" "M,D,d,U,r,R,D,d,r,r,I,Y,i,T,*A,*r"))]
"xtensa_valid_move (SImode, operands)"
"@
movi.n\t%0, %x1
mov.n\t%0, %1
mov.n\t%0, %1
+ %v1l32i\t%0, %1
+ %v0s32i\t%1, %0
%v1l32i.n\t%0, %1
%v0s32i.n\t%1, %0
%v0s32i.n\t%1, %0
@@ -1286,13 +1288,11 @@
movi\t%0, %1
const16\t%0, %t1\;const16\t%0, %b1
%v1l32r\t%0, %1
- %v1l32i\t%0, %1
- %v0s32i\t%1, %0
rsr\t%0, ACCLO
wsr\t%1, ACCLO"
- [(set_attr "type" "move,move,move,load,store,store,move,move,move,move,move,load,load,store,rsr,wsr")
+ [(set_attr "type" "move,move,move,load,store,load,store,store,move,move,move,move,move,load,rsr,wsr")
(set_attr "mode" "SI")
- (set_attr "length" "2,2,2,2,2,2,3,3,3,3,6,3,3,3,3,3")])
+ (set_attr "length" "2,2,2,3,3,2,2,2,3,3,3,3,6,3,3,3")])
(define_split
[(set (match_operand:SHI 0 "register_operand")
--
2.39.2
^ permalink raw reply [flat|nested] only message in thread
only message in thread, other threads:[~2024-03-14 17:31 UTC | newest]
Thread overview: (only message) (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2024-03-14 17:30 [PATCH] gcc: xtensa: reorder movsi_internal patterns for better code generation during LRA Max Filippov
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).